Near singular Matrix when extending sample size

[altaus]

Hi,

I have set up a program in which I use a 3sls system to forecast bond yields. First, I set up the equation for the system (without estimating). Then, I go on to create the variable forecasts through ARMA models, and by importing manually inputted values from Excel. This later steps seems to be causing the problem in the estimation step (Error "Near singluar matrix"), which is the next one, but this only happens when I use the latest row of data (August data in the Excel file). When I remove the August data it works fine. I have also noticed that if I reduce the estimation sample by starting it from @first +3, it does work but provides significantly different results to those with the data up to July.

I would appreciate if you could have a look. I imagine I am overlooking something, but I can't make up my mine.

Thanks a lot,
Marta

P.D The program requires the path link to the Excel file twice, and in pop-up box you'll need to select SR.

[altaus]

Hi,

Having read posts about the same issue, I have checked the coefficients in C and I have realized that the intercept for japan gets really messed up when adding the last data point (this is when the error occurs). I can't figure out why but if this is a collinerarity problem, I don't understand why just one data point would make such a difference for the series. If I substitute the new data point in jp_lr and jp_sr for the previous one, I don't get the error message anymore. Any idea why this is happening?

The C(7) coefficient is around 1.4 level in most of the iterations, but when using the data for August it goes up to -228978.174. This also seems to have an impact on the German intercept C(1) which is around 0.7 for most of the monthly iteration, except for July when it goes down to 0.3 and August when it goes up to 9.73.

Thanks for your help in advance.

Marta

[EViews Gareth]

You'll probably receive more help if you just provide an EViews workfile with the system (I presume?) object in it, then tell people which sample works and which doesn't.

[altaus]

Ok, I attach two workfiles with the results here

1. g4 dataset NOT WORKING >> if you try to estimate the system it provides with the near singular error.
2. g4_dataset OK substituting values>> the last observation in jp_lr and jp_sr has been changed and the system works

[EViews Gareth]

Neither system estimates for me.

[altaus]

I don't know what to say. I have tried them again by downloading them in my computer and running both of them and I get the "near singular matrix" in the first. the second one runs normally. What error do you get?

[EViews Gareth]

Near singular matrix. I would guess that both are borderline/knife-edge singular, and machine slop is what is causing it to fall one side or the other.

[altaus]

so what do you recommend for this? It is Japan that is causing the problem, but with the extra data point. Does that make sense to you?